Large language models and the IndiaAI Mission
A large language model is generative AI trained on vast text to predict the next token, which lets it write and answer but also invent facts. The IndiaAI Mission, approved on 7 March 2024 with ₹10,371.92 crore, funds compute and home-grown foundation models; MeitY reported its progress to the Lok Sabha on 5 August 2026.
Syllabus: Prelims · Current Affairs & IR; GS2 · Government Policies; GS3 · Science & Technology
Why this matters for UPSC
GS3 science and technology and government policy: Prelims can ask how a large language model works, the IndiaAI Mission's pillars and who implements it; Mains, India's path to sovereign AI and the risks of generative AI.
In short
A large language model is generative AI trained on vast text to predict the next token; the IndiaAI Mission (2024) funds the compute, data and models India needs.
India's AI push rests on home-grown models and shared compute; it must cut GPU dependence and check hallucination before relying on models in public services.
Why it matters
A farmer asking a chatbot about crop disease in Marathi, or a clerk drafting a reply in Tamil, depends on a model that understands Indian languages. Many widely used models are built abroad, and India's AI compute still runs on imported GPUs. Home-grown models and shared computers let Indian startups, researchers and departments build such tools without renting everything from overseas. The same technology can mislead: a confident wrong answer about a medicine or a law can harm the person who trusts it. That is why the mission also pays for work on bias, privacy and auditing, and why the global summit stressed trusted and energy-efficient AI.
Prelims pointers
- IndiaAI Mission's seven pillars: Compute Capacity, Innovation Centre, Datasets Platform, Application Development Initiative, FutureSkills, Startup Financing, and Safe & Trusted AI.
- The compute pillar plans AI compute of 10,000 or more GPUs built through public-private partnership, plus an AI marketplace offering AI as a service and pre-trained models.
- IndiaAI FutureSkills sets up Data and AI Labs in Tier 2 and Tier 3 cities and expands AI courses at undergraduate, master's and PhD level.
- The AI Impact Summit 2026's seven Chakras include Democratizing AI Resources, AI for Science and Human Capital Development; PIB called it the first global AI summit in the Global South.
- The National Mission on Interdisciplinary Cyber-Physical Systems, run by the Department of Science and Technology, has set up 25 Technology Innovation Hubs at academic institutions.
- The Semicon India Programme, launched in January 2022, supports semiconductor fabs, display fabs, packaging units and Indian-designed chips.
Practise a UPSC-style question
With reference to large language models, consider the following statements:
1. They generate text by predicting the next token from statistical patterns learnt in training.
2. Their parameters are copies of the training documents, stored word for word so they can be looked up.
Which of the statements given above is/are correct?
- (a) 1 only
- (b) 2 only
- (c) Both 1 and 2
- (d) Neither 1 nor 2
Show the answer
Answer: (a) 1 only
Answer: (a) 1 only. A large language model predicts tokens; it does not look up stored copies of documents.
Statement 1: correct. As NIST's Generative AI Profile (2024) explains, large language models predict the next token or word, approximating the statistical distribution of their training data.
Statement 2: incorrect. NIST explains that generative models produce outputs approximating the statistical distribution of their training data; they predict text rather than retrieve stored documents word for word.
Eliminate: Statement 1 is how every large language model generates text, so it is correct. Out: (b), (d).
Eliminate: Statement 2 mistakes a predictive model for a stored library, so it is incorrect. Out: (c).
Trap: Sounds right. A model that quotes facts seems to look them up, but it only predicts likely tokens, which is also why it can hallucinate.
Remember: A model's output is predicted token by token, not looked up.
More Explainers
- State Legislative Councils: Articles 168-172 and the graduates' and teachers' seats Polity
- Whales and dolphins in Indian waters: legal protection, IUCN status and strandings Environment
- The Aligarh Movement, AMU and the Article 30 minority test Polity
- Environmental clearance by States: SEIAA, SEAC and the EIA Notification, 2006 Environment
- Software vulnerabilities: CVE, CVSS, coordinated disclosure and CERT-In's role Sci-Tech
- Drug trafficking law in India: NDPS Act, PIT-NDPS, NCB and NCORD National
- Choosing the UN Secretary-General: Article 97, UNSC veto and straw polls World & IR
- UN Convention against Torture, and why India has not ratified it Polity
- NavIC: why India's navigation constellation cannot give a standalone position Sci-Tech
- International Energy Agency: membership, 90-day oil stocks and India's status World & IR
- PM Surya Ghar: Muft Bijli Yojana and the rooftop solar subsidy Economy
- HFSS foods: FSSAI's draft limits and front-of-pack warning labels National